How to find AI use cases in your own business

A broker I work with in London had a colleague whose entire day was one job. Open a property listing. Find out which company was behind the development. Look up the directors. Hunt for an email address. Check it was not dead. Type all of it into the CRM. Then do it again.

Nobody in that business ever asked which AI tool to buy. They knew exactly where the time was going, because a person was standing in it.

That is the whole problem with how most people look for AI use cases. They sign up for something first, then go hunting for a job to give it.

Start with the work, not the tool

The question is not "what can this tool do". It is "what are we already doing every week that nobody enjoys".

Three words tell you if a job qualifies.

Repetitive. You do it often, and roughly the same way each time. Entering the same data. Building the same weekly report. Uploading the same kind of file to the same folder.

Rule based. You can describe the rules out loud. If the invoice is under fifty dollars, approve it. If the customer is in the UK, route it to the UK rep. Plenty of jobs are repetitive but need a fresh judgement every time, and those are much harder to hand over.

Routine. No real creative problem solving. Deadline reminders. Survey sends. Product listings. Recurring notifications.

Here is the part most people get backwards.

If a job passes all three, you probably do not need AI at all. Plain automation is faster, cheaper and completely predictable. Using a language model to make a decision you could write as a rule is a worse system that costs more and occasionally invents an answer.

AI earns its place where the input is messy. Not where the logic is simple.

Then decide what to build first

Once you have a list of candidates, you need an order. Two questions per job: how long does it take once, and how often does it happen.

Most people build the thing that irritated them most this morning. That is almost never the thing worth building.

Small wins are the underrated box. A five minute job that happens ten times a day is not costing you fifty minutes. It is costing you fifty minutes plus ten interruptions, and the interruptions are the expensive part.

Do the maths before you build anything

Take a handover process. New customer signs, and someone moves them across to the team who will actually deliver. Say it takes 45 minutes.

Forty five minutes is nothing. Nobody escalates a 45 minute job.

Now count them. Twelve a week is about 40 hours a month. You are spending a full working week, every month, on a job nobody thinks is a problem.

Say the first version takes 16 hours to build and covers 80% of it. That pays for itself inside the first month and keeps paying every month after.

You might be thinking: fine, but my version saves an hour a month, not forty. Is that worth building?

Usually not on its own. But you will find six of them, and they compound. The reason to write the numbers down is that it stops you building the exciting thing and skipping the boring one that was quietly eating a week a month.

Two places nobody looks

Inside the steps, not the process. Write out what actually happens when a lead signs up. They fill in the form. They get a confirmation. Someone assigns them. Someone reads the CRM. Someone writes an email. A meeting gets booked.

Do not try to automate that. Automate one step of it. Enrich the lead so the rep is not googling the company. Route it so nobody has to decide who takes it. Ship that, watch it run, then take the next step. Whole journeys fail as projects. Single steps ship in a week.

The work you are not doing. This is the one people miss entirely. There are jobs you skip because there is no time. Nobody checks every competitor's site each week. Nobody writes up every meeting properly.

So the better question is not "where does AI replace me". It is "what would I do if it cost almost nothing to do it". That is where the interesting use cases live, because nothing is being taken away from anyone.

What AI is actually good at

In production, models are reliably good at four things. Not forty.

Invoices are the clearest example. Every company's invoice looks different, so you cannot say "read the number in the top right". A model can read the document and pull out the sender, the number, the date, the total and the currency, and hand you clean fields.

Same shape works on meeting notes, customer feedback, CVs and contracts. Anything where the information you need is trapped in a document written by someone who was not thinking about your database.

The one pattern worth learning first

Most systems are made of a few repeating shapes. This is the one you will use most.

Something arrives, something decides, it goes to the right place. Leads to reps. Tickets to teams. CVs to roles. Invoices to whoever handles exceptions.

Use rules where the answer is objective, like company size or country. Use a model where it is not, like working out what a customer email is actually asking for. When you have a real rule, use the rule. You will always understand why a rule did what it did.

The other shapes worth knowing: a plain automation moving data between two tools, an automation with one AI step in the middle, a scheduled report pushed into Slack instead of sitting in a dashboard nobody opens, and an agent that pulls together context a person would otherwise reconstruct across six tabs.

Almost everything useful is a few of those stacked together.

Ship the rough one

Build for the volume you have now.

An agent that reads forty submissions a day and posts a summary might fall over at a thousand. That is fine. It will have told you whether anyone actually reads the summary, which is the thing you needed to know. When the volume arrives, replace the fragile part with something deterministic that counts and sorts properly.

If you wait until it is right, you never find out whether it was worth building.

The goal was never to automate everything. It is to stop people spending their week on work that does not need a person in it.

If you want to see what this looks like finished, the lead pipeline for that London broker is the job I opened with, and the rest of the builds are written up the same way.

← All articles